Submesh allocation in 2D-mesh multicomputers: partitioning at the longest dimension of requests

Two adaptive non-contiguous allocation strategies for 2D-mesh multicomputers are proposed in this paper. The first is first-fit-based and the second is best-fit-based. That is; for a given request, the proposed first-fit-based approach tries to find a free submesh using the well-known first-fit strategy, if it fails, the request at hand is partitioned into two sub-requests that are independently allocated using the first-fit approach. Partitioning is gradually performed at the longest dimension of the parallel request. This partitioning mechanism aims at (i) lifting the condition of contiguity, and (ii) at the same time maintaining good level of contiguity. Gradual partitioning of requests produce two sub-requests one of which is relatively big and as close as possible to the square-shape and, thus; reducing communication latency caused by non-contiguity. Using extensive simulations, we evaluated the proposed strategies and compared them with previous contiguous and non-contiguous strategies. Simulation outcomes clearly show the proposed allocation schemes produce the best Average Response Time (ART), the Average System Utilization (ASU) and also produce relatively low communication overhead.